ECTI Transactions on Computer and Information Technology (ECTI-CIT)

AquaLink in HAB Detection: Integrating IoT and 3D-Printed PETG for Monitoring Aquaculture Conditions Conducive to HAB

2025-08-07T09:29:32+07:00

Harmful algal blooms (HABs) pose a serious threat to aquaculture and environmental health, often resulting in considerable ecological and economic impacts. Conventional water quality monitoring techniques, often manual and time-consuming, are inadequate for the timely detection of conditions that promote HAB formation. To overcome these limitations, the AquaLink system was developed by integrating the Internet of Things (IoT) technology with 3D-printed polyethylene terephthalate glycol (PETG) enclosures, enabling scalable, real-time, and cost-effective monitoring of water quality. The system employs sensors to measure essential parameters, including atmospheric pressure, temperature, and turbidity, with data transferred through Raspberry Pi and ESP32 controllers to an IoT dashboard for real-time analysis and visualisation. PETG-based casings were combined with IoT-enabled sensors to improve durability and reduce biofouling in aquatic environments. Prototypes were tested across different water bodies to validate performance under real-world conditions. The results demonstrated that the system effectively provided real-time monitoring of aquaculture environments, allowing the early identification of HAB risks through continuous tracking of water quality indicators. Beyond its technical contributions, AquaLink offers societal benefits by serving as a low-cost, efficient tool that reduces sh mortality, limits environmental degradation, and enhances food security. The flexibility and scalability of the system make it applicable to small-scale and industrial aquaculture operations, fostering sustainable practices through advanced environmental monitoring.

Assessment of Factors Influencing IoT Adoption in Jordanian Healthcare Using AHP Technique

2025-08-14T11:14:24+07:00

The Internet of Things (IoT) has emerged as a transformative technology with vast potential, particularly in the healthcare sector, where it can significantly enhance operational efficiency and improve the quality of care. However, IoT adoption in healthcare, especially within Jordanian public hospitals, remains limited and is still in its early stages, with numerous challenges impeding widespread implementation. The present study seeks to provide deeper insights into the critical factors influencing IoT adoption within this context. Data were collected through a structured survey administered to a panel of experts from Jordanian public hospitals with substantial knowledge and experience in IoT technologies. Utilizing the Analytic Hierarchy Process (AHP), the study calculates the relative importance of factors categorized within four dimensions using an integrated TOE and HOT-Fit framework. The findings highlight Top Management Support (0.238), Relative Advantage (0.223), Compatibility (0.146), and Government Support (0.134) as the most influential criteria for IoT readiness. The study provides enriched theoretical perspectives and practical guidance for policymakers and hospital administrators seeking to enhance institutional readiness and promote IoT adoption in healthcare.

Optimizing Cloud-Integrated Computer Networks: Strategies for Enhanced Performance and Security

2025-08-21T10:45:18+07:00

This study proposes a unified, Python-based simulation framework to enhance the performance and security of cloud-integrated computer networks. The framework concurrently addresses two critical aspects, load balancing and intrusion detection, within a single reproducible environment. Using the CICIDS2017 dataset, a Random Forest classifier was used to detect a wide range of network attacks with high accuracy. To simulate realistic traffic behavior, synthetic data were generated for performance metrics such as latency, throughput, and packet loss. Load balancing is evaluated using round-robin and random assignment strategies across virtual servers, illustrating the trade-offs between uniformity and randomness in the request distribution. The experimental results demonstrated a classification accuracy of 99.79%, with precision and recall metrics supporting the robustness of the selected model. Feature importance analysis highlights the key indicators of anomalous traffic, and confusion matrices and precision-recall curves validate the detection performance. Additionally, the simulated network KPIs provide a scalable approximation of the Quality of Service under varying load scenarios. The proposed research is the only one that suggests an integrated and data-driven approach to fill the gap documented in previous studies, where network security and resource optimization are frequently studied independently. The proposed framework provides a convenient basis for additional scholarly and industrial investigations in the field of secure cloud networking.

Optimizing Crop Yield Predictions through Satellite Data Fusion and Machine Learning

2025-08-31T10:45:29+07:00

Accurate crop yield estimation is crucial for sustainable agriculture and food security, especially in Maharashtra, where climate variability significantly impacts crop growth. This study utilizes satellite data from MODIS, Landsat, Sentinel-1, and Sentinel-2 to predict the yields of 22 crops across 36 districts. Machine learning models, including Random Forest, Gradient Boosting, and SVM, were evaluated using RMSE, MAE, and R2 metrics. Random Forest outperformed the others, achieving R2 values above 0.70 for all crops, with a peak R2 of 0.93. Incorporating seasonal and permuted feature data further enhanced predictions, demonstrating the efficacy of integrating satellite data and machine learning for agriculture. Keywords: Machine learning, MODIS, Landsat-8, Sentinel-2, Sentinel-1, crop yield, features, vegetation indices.

Optimized IoT-Based Multimodal Fusion for Early Forest Fire Detection and Prediction

2025-07-21T09:07:33+07:00

Forest and land fires are recurring ecological disasters that pose serious threats to environmental sustainability, particularly in vulnerable regions like Indonesia. Conventional fire detection methods using only visual or single-sensor data often suffer from low accuracy in poor lighting, thin smoke, or extreme weather. This study proposes an IoT-based multimodal system that combines visual imagery and real-time meteorological sensor data. Fire detection was conducted using the YOLOv11 model, trained for 50 epochs with the SGD optimizer. The model achieved a precision of 87.9%, recall of 79.7%, mAP@0.5 of 87.7%, and mAP@0.5:0.95 of 53.7%. Detected images are further classified using a hybrid ViT-GRU model, which achieves 99.97% accuracy by capturing spatial and temporal fire patterns. We performed fire detection using an LSTM model optimized with Optuna and SMOTE, yielding 92.66% accuracy and an AUC of 1.00. The decision-level fusion approach integrates visual and sensor outputs to improve the accuracy and contextual relevance of the nal prediction. We deployed the system in a real-time Streamlit dashboard connected to cloud-based data acquisition. Results show that this multimodal approach significantly improves the reliability of early re detection and risk prediction.

Hybrid Emotion Classification of MOOC Reviews Using the NRC Lexicon and a Multi-Channel Deep Learning Model

2025-09-18T11:19:28+07:00

Text-based emotion recognition has received extensive attention in applied computing research, but its effectiveness in online learning contexts remains limited. In this study, we introduce the TriFusion Attention Network, a hybrid deep learning model that classies emotions in Massive Open Online Course (MOOC) reviews. Using the NRC Emotion Lexicon, we annotated learner reviews and designed the model to integrate multiple channels capturing both semantic and affective information. Its architecture combines Bidirectional Long Short-Term Memory (BiLSTM), Bidirectional Gated Recurrent Units (BiGRU), Convolutional Neural Networks (CNN), and attention mechanisms to model the complexity of learner feedback effectively. Experiments conducted on Coursera reviews demonstrate that the model effectively identifies both explicit and subtle emotional cues, achieving over 95% accuracy, F1-scores around 0.95, and AUC-ROC values approaching 0.99 on both balanced and imbalanced datasets. These results confirm that the proposed approach achieves superior performance compared to existing methods and facilitates improved learner engagement while offering richer analytical insights into their experiences.

AI-Driven Sign Language Recognition System with NLP-Enhanced Transcription

2025-08-31T10:57:26+07:00

Sign language is a critical communication medium for deaf and hard-of-hearing individuals, yet the diversity of over 7,000 sign languages worldwide presents significant challenges for automated recognition systems. This paper presents a novel approach to sign language recognition (SLR) that integrates computer vision techniques with advanced natural language processing (NLP) to improve transcription accuracy and contextual relevance. Our system employs a two-stage architecture: first, a gesture recognition component utilizing MediaPipe Holistic for landmark extraction and Long Short-Term Memory (LSTM) networks for classification; second, a text enhancement module using bidirectional LSTM for contextual correction and grammatical improvement. Experimental results demonstrate that our NLP-enhanced system achieves 98.46% accuracy in gesture recognition while significantly improving the grammatical correctness and contextual coherence of the generated text compared to systems without NLP enhancement. The system can successfully identify missing function words, add appropriate punctuation, and correct grammatical errors in real-time. While primarily focused on American Sign Language (ASL), our approach provides valuable insights for developing more effective and inclusive SLR technologies for various sign languages. These advancements represent a meaningful step toward bridging communication gaps between signing and non-signing individuals, potentially enhancing accessibility in educational, professional, and social environments.

A Hybrid Transformer-Based Deep Neural Network for DDoS Detection: A Comparative Evaluation Across Modern Architectures

2025-10-09T09:23:52+07:00

DDoS attacks remain a major threat to network infrastructures. While deep learning is applied for detection, prior studies often lack standardized comparison and stability evaluation under consistent settings. This study systematically evaluates nine deep learning models including MLP, CNNs, ResNet1D, and attention-augmented architectures under consistent experimental settings and introduces two novel models: TDNN (Transformer-based) and ATDNN (attention-enhanced) for capturing complex traffic patterns. Using a balanced real-world dataset, all models were trained over five independent runs, with performance assessed via accuracy, precision, recall, and F1-score. TDNN achieved the highest performance (Accuracy: 0.9653 ±0.0018; Precision: 0.9659 ±0.0016; Recall: 0.9653 ±0.0018; F1-score: 0.9653 ±0.0018), while simpler models such as DNN, MLP, and LSTMClassier also performed competitively with lower variance. The study further analyzes learning behaviors and evaluates deployment potential, highlighting that well-tuned deep learning models, particularly TDNN, can support real-time DDoS detection in enterprise and edge computing environments.

Fuzzy Inference Approach for PM₂.₅ Modelling with High Accuracy and Low Complexity

2025-10-09T09:31:09+07:00

PM_2.5is a silent yet severe pollutant that accumulates in the human body, causing long-term health issues such as lung cancer. This study proposes a novel fuzzy inference system (FIS) for PM_2.5forecasting, addressing the nonlinear and dynamic nature of air pollution. Unlike complex, data-intensive black-box models, the proposed FIS is transparent, interpretable, and simple to implement. It uses only two lagged PM_2.5change rates and nine fuzzy rules for accurate prediction. The model requires no geographical or emission-source data, which are often costly and region-specific. Fuzzy rules are derived from natural PM_2.5rise-and-fall patterns, ensuring logical consistency and minimal inputs. Using data from Chiang Mai, Thailand —one of the most polluted cities —the model was benchmarked against MLR, MLP, LSTM, SVM, and Gradient Boosting. The FIS achieved up to 5% higher accuracy. Although the Diebold-Mariano test found no significant difference, FIS showed comparable robustness with 49% fewer parameters and 56% fewer FLOPs. Optimal performance occurred at three input lags and 27 fuzzy rules, balancing accuracy and complexity. Moreover, the Chiang Mai FIS generalized well to other PM_2.5-affected cities —Bangkok, Jakarta, and Ho Chi Minh City—without modifications, and maintained reliability for both daily and extended hourly forecasts.

Enhanced Load-Aware Handover Algorithm for High-Density IEEE 802.11 ESS Networks

2025-08-02T18:38:03+07:00

High-density Wi-Fi networks, particularly in Extended Service Set (ESS) environments, frequently experience performance degradation due to suboptimal handover mechanisms that rely exclusively on Received Signal Strength Indicator (RSSI). Such approaches often lead to traffic imbalance, increased packet loss, and reduced Quality of Service (QoS). This paper introduces a novel implementation of a load-aware handover strategy that integrates both RSSI and real-time access point (AP) load metrics to optimize handover decisions. Through its adaptive weighting function and penalty mechanism, ELAHA dynamically balances signal strength with AP load conditions, achieving improved network efficiency and user experience in dense deployment scenarios. Results demonstrate that ELAHA significantly outperforms the conventional RSSI-Based Algorithm (RBA), achieving lower latency, reduced jitter and packet loss, decreased handover frequency, and enhanced overall throughput. These findings highlight ELAHA's potential as a robust and scalable solution for maintaining consistent QoS in high-density Wi-Fi networks.

Awareness of Check-In Patterns for an Adaptive Framework in Next POI Recommendation

2025-08-17T15:54:13+07:00

The recommendation of the Next Point-of-Interest (Next POI) is pivotal in the domain of location-based services, as it forecasts a user's subsequent check-in based on their historical movement patterns. Although prior researches have recognized and acknowledged the diversity in individual travel behavior, the methodologies for effectively distinguishing these patterns remain somewhat ambiguous and unclear. This particular challenge becomes more complex when users engage in check-ins at irregular locations or times, which consequently complicates the process of forecasting the Next POI. To address this issue, we aim to analyze the check-in patterns to improve the Next POI recommendation process. To address this particular concern, we analyze check-in patterns to improve the Next POI recommendation process. We propose AFNextPOI (Awareness of Check-In Patterns for an Adaptive Framework in Next POI Recommendation), which enhances check-in pattern analysis through pattern-based features. This research implements strict privacy protection measures, utilizing only anonymized check-in data to ensure no user profile information is accessible or analyzed. Experiments conducted on two real-world datasets demonstrate that the AFNextPOI framework achieves superior performance compared to state-of-the-art models in terms of Recall and NDCG metrics, thereby validating the effectiveness of our approach.

UAV-Based mmWave MIMO with PSO-Enhanced Hybrid OMA-NOMA Access

2025-10-10T10:54:03+07:00

This paper presents a UAV-based communication system that integrates hybrid Orthogonal Multiple Access (OMA) and Non-Orthogonal Multiple Access (NOMA) schemes, operating in the millimeter-wave (mm-Wave) frequency band and supported by Multiple-Input Multiple-Output (MIMO) technology. The proposed Hybrid OMA/NOMA-mm-Wave MIMO framework is designed to enhance overall system performance by delivering high-capacity wireless connectivity to ground users (GUs). UAVs act as aerial base stations (BSs), offering rapid and flexible communication services in diverse real-world scenarios, including natural disasters, areas lacking fixed infrastructure, and temporary coverage at large public events. To improve NOMA's performance, a Particle Swarm Optimization (PSO) algorithm is employed for optimizing power allocation (PA), ensuring fairness between near and far users. Furthermore, a user pairing mechanism integrated with optimized power allocation is introduced to enhance the UAV-BS performance in mm-Wave-NOMA scenarios. The channel model considers both line-of-sight (LoS) and non-line-of-sight (NLoS) conditions, incorporating angle of departure (AoD), angle of arrival (AoA), and Doppler effects. Simulation results demonstrate that NOMA outperforms OMA in specific scenarios, while OMA remains more effective in others. PSO-based power allocation significantly surpasses fixed PA schemes in NOMA systems.

Scalable Influence Maximization Using Ant Colony Optimization with Attribute-Based Scouting

2025-10-09T09:15:18+07:00

Influence Maximization (IM) is a vital problem in social network analysis that focuses on identifying a small subset of influential nodes to maximize the spread of information across a network. Traditional influence maximization algorithms, including greedy and heuristic-based methods, often struggle with scalability and efficiency, especially when applied to large-scale networks. To overcome these limitations, we propose a novel Hybrid Ant Colony Optimization (HybridACO) algorithm that integrates a neighbor scouting strategy based on attribute similarity. This approach utilizes the inherent network structure by combining the global search capability of Ant Colony Optimization (ACO) with a local scouting mechanism that selects nodes based on their neighbors' influence potential and attribute similarity. By integrating attribute-driven scouting, HybridACO ensures that the selected nodes are not only topologically influential but also contextually relevant for the diffusion process. Comprehensive evaluations on both synthetic and real-world benchmark networks show that the proposed algorithm significantly surpasses existing state-of-the-art (SOTA) methods in influence spread, computational efficiency, and robustness.

TOEHBCHAIN: A Framework for Blockchain-Based Food Distribution Transparency

2025-10-23T17:13:58+07:00

This study develops and validates TOEHBCHAIN, a readiness-driven blockchain adoption framework designed to enhance transparency, efficiency, and accountability in Indonesia's Makan Bergizi Gratis (MBG) program. Using a mixed-methods design, data were collected from 323 student beneficiaries, 23 schools, and qualitative interviews with government officials, vendors, and administrators. The TechnologyOrganisationEnvironmentHuman (TOEH) framework guided readiness assessment, while blockchain architecture was structured into four layers: Application, Smart Contract, Data, and Integration. Triangulation through expert validation and Focus Group Discussions strengthened reliability. Results show high organisational and environmental readiness but moderate technological understanding and limited human competence. Qualitative findings emphasise the need for automated reporting, real-time tracking, and inclusive interfaces. The TOEHBCHAIN framework integrates TOEH dimensions with blockchain features such as RBAC security, lightweight PWA design, and interoperability with government databases. Validation confirmed its feasibility, contextual relevance, and replicability. TOEHBCHAIN effectively aligns stakeholder readiness with blockchain functionalities to address digital transparency gaps in food distribution, offering an adaptable model for broader public welfare applications.

Energy-Efficient Hybrid Learning for Secure Wireless Sensor Networks

2025-10-23T11:14:41+07:00

Wireless Sensor Networks (WSNs) power critical applications from environmental monitoring to Internet-of-Medical-Things healthcare yet their tiny batteries and low-end microcontrollers leave them exposed to network-layer Denial-of-Service (DoS) attacks such as Blackhole, Grayhole, Flooding and TDMA scheduling. Signature IDSs miss zero-day variants and shallow machine-learning detectors produce many false alarms, while running monolithic deep-learning models on every node exhaust energy reserve. We introduce a two-stage hybrid IDS in which each sensor executes an integer-only rule filter that costs ≤0.05 mJ per packet and discards ≈95% of benign traffic, forwarding only flagged flows over BLE/LoRa to an edge gateway. There, a 50 %-pruned, 8-bit CNN-LSTM processes 32-window batches in 28 mJ and ≈42 ms. Experiments on the public WSN-DS corpus, augmented by ns-3 simulations of a 50-node LoRa network, show that the scheme achieves 98 % accuracy, 0.93 macro-F1 and minority-class recalls of 0.840.95 while extending network lifetime (T50) to 69 days an 82 % gain over on-node GRU and 35 % over a signature IDS. Removing the rule filter erases most of the lifetime benefit without affecting accuracy, confirming that local triage, not downsized deep models, is the key to energy efficiency. The evaluation answers four research questions covering optimal hybrid architecture, rule-filter tuning, node-level energy overhead, and performance trade-offs against traditional ML and standalone DL baselines. These findings demonstrate that intelligent workload partitioning can deliver deep-learning-level security without shortening the lifetime of resource-constrained WSN deployments.

Detecting Manipulation in NFT Market Using Graph-based Deep Learning

2025-10-23T10:09:27+07:00

The rise of non-fungible tokens (NFTs) has increased the risk of fraud and market manipulation. This study introduces a method for detecting wash trading in the NFT marketplace using Graph Neural Networks (GNNs) applied to Ethereum blockchain transaction data. We constructed a heterogeneous graph, used Depth-First Search for labelling, and extracted graph features, including PageRank and degree centrality. We evaluate various classification models: Multilayer Perceptron (MLP), Graph Convolutional Neural Network (GCN), and Heterogeneous Graph Convolutional Neural Network (HeteroGCN). The results show that GNN models, particularly the feature-enhanced HeteroGCN, exhibit superior performance compared to featureless models and traditional tabular baselines. The key contribution of this study is that PageRank and Degree Centrality features significantly improve the accuracy of identifying transactions involved in market manipulation.

Comparative Analysis of GLCM Features for ALL Detection using Convolutional Neural Networks

2025-10-23T10:41:00+07:00

Acute Lymphocytic Leukemia (ALL) is the most prevalent childhood cancer. Early diagnosis of ALL is essential, since timely and appropriate therapy significantly improves survival rates. ALL entails abnormal white blood cell (WBC) proliferation that weakens the immune system. Despite numerous studies on ALL identification using microscopic blood smear images, challenges remain due to the structural complexity of blood cells, image noise, intensity inhomogeneity, and cell overlap. Deep learning techniques, particularly CNNs, have shown high efficacy in medical image analysis. However, some studies employ GLCM features that lack an empirical basis for their selection. The study aims systematically evaluate nine distinct GLCM features to provide evidence-based reference for their use in ALL detection. Foundational CNN architecture was used as a classifier to establish a performance baseline with raw color images. Then we compared the baseline to models enhanced with each of the nine GLCM features using the public C-NMC 2019 dataset. The results revealed that GLCM Dissimilarity was the most effective feature, yielding an accuracy of 94.37% and an F1-score of 0.9437. The GLCM Dissimilarity significantly outperformed other GLCM features, followed by GLCM Entropy (93.84%), GLCM Mean (93.57%), GLCM Contrast (93.55%), GLCM Standard Deviation (93.13%), and GLCM ASM (90.98%). A model using only color images achieves an accuracy of 90.53%. It is important to note that although several GLCM features improved performance, among the nine features evaluated, those not listed among the top performers achieved lower accuracy than the color image baseline, highlighting the importance of careful feature selection. These results indicate that specific GLCM features, particularly GLCM Dissimilarity, can substantially enhance CNN-based ALL classification, underscoring its potential as a robust image descriptor for diagnostic applications.

Enhancing Graph-Based Sentiment Analysis Models with Hill Climbing

2025-10-17T11:18:37+07:00

In the current study, sentiment graphs were constructed in which the nodes represented emotion-laden words, and the edges depicted their weighted semantic associations. To improve the model, the hill climbing method was employed, which iteratively adjusted parameters to achieve increasingly higher classification accuracy. The developed system employed a combination of graph neural networks (GNNs) and hill climb-based optimisation to improve the efficiency of sentiment categorisation. The experiment's outcomes reveal that the suggested model reached a maximum accuracy of 96.95%, which is higher than traditional sentiment analysis methods and thus proves its appropriateness for emotion-aware text representation. The experimental findings confirm that GNN-based sentiment representation and hill climbing optimisation effectively leverage the intricate emotional relationships, resulting in better sentiment classification. The graphs illustrating optimisation progress and the structure of the sentiment graph further demonstrate the effectiveness of our method.

Hybrid Deep Learning and Machine Learning Framework for Automated Tomato Leaf Disease Classification

2025-10-23T10:35:31+07:00

Tomato leaf diseases significantly impact crop productivity, necessitating accurate and efficient diagnostic tools. This study proposes a hybrid framework that integrates deep learning-based localization and segmentation with handcrafted feature extraction and classical machine learning for tomato leaf disease classification. Specifically, YOLOv8 is used for object detection and SAM for segmenting diseased regions. Features are then extracted using HSV color space, GLCM, and LBP descriptors. To address class imbalance, the SMOTE technique was applied, expanding the original 48,243 image dataset to 102,465 balanced samples across 11 disease categories. Multiple classifiers were evaluated, with Random Forest achieving the highest performance over 90% accuracy and a macro F1-score of 0.90. Importantly, recall for minority classes improved markedly after balancing. The proposed system demonstrates strong potential for deployment in real-world agricultural environments due to its low computational cost and robustness under varying conditions. Future work will explore multi-crop generalization, real-time inference, and eld validation under challenging conditions such as lighting variation and occlusion.

A Novel Approach to Dairy Sales Forecasting: Multi-Perspective Fusion Bi-LSTM Coupled with Universal Scale CNN

2025-09-19T08:04:10+07:00

In today's world, accurately predicting sales is important for minimising costs and improving overall profits. A wide range of understanding of organisational performance and future sales can be accomplished through improved customer service strategies. It aids in enhancing product returns and lowering lost sales, leading to more efficient production planning. The sales prediction in dairy products reflects distinctive challenges, predominantly due to the quality of these products, which is closely connected to consumers' health. To overcome the problem, the proposed research em- ploys an effective DL (Deep Learning) based technique for forecasting the sales of dairy products by analysing the dairy goods sales dataset from an openly available website. The proposed research utilises Universal Scale CNN (Convolutional Neural Network), a 1D CNN, which is capable of learning the features at optimal and effective rates. The following features are passed to the Multi-Perspective based Bi-LSTM (Bidirectional Long Short-Term Memory), which is capable of learning features effectively by reducing error rates in predicting sales rates of dairy-based products. The overall performance of the proposed Multi-Perspective Fusion Bi-LSTM with Universal Scale CNN is evaluated with performance metrics, including RMSE (Root Mean Squared Error), MAE (Mean Absolute Error) and MSE (Mean Squared Error). These performance metrics evaluate the proposed model's effectiveness in forecasting dairy product sales.